Local timeline

The paper introduces ARBOR, a framework that tracks reasoning states in long chains of thought to locate failures.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.22717
#MachineLearning

Causal Edit Serialization introduces a framework that expresses revisions within the same causal token stream using special pointer tokens paired with predicted source addresses.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.22820
#MachineLearning

Recent studies show position bias remains a critical vulnerability in large language models. Increasing model scale and Chain-of-Thought prompting significantly mitigate this issue.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.22807
#MachineLearning

MambaAct swaps the transformer action expert in π0.5 for a compact Mamba-based expert and adds a Context Compressor for visual-language conditioning.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.22951
#Robotics

PASS-VLN introduces a passability-aware local-command vision-and-language navigation benchmark and model.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.22812
#Robotics

Lens Technology is expanding into AI server infrastructure through a proposed US$25 million acquisition of a controlling stake in PMG International.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261019VL208/ai-server-manufacturing-cooling-acquisition-infrastructure.html
#AI

Taiwan has launched its first proof-of-concept project to shift renewable electricity using energy storage.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261019PD206/taiwan-electricity-testing-taiwan-power-company-capacity.html

AnchorQ uses flow reversal and value-guided sampling for coarse-to-fine policy improvement.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.22852
#Robotics

GPT-6-Astra demonstrates competence in robotic manipulation and navigation.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.22471
#Robotics

Anthropic signed its first data center lease in Canada at the Point Tupper campus planned by EverWind Fuels in Nova Scotia.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261019PD228/anthropic-data-center-canada-capacity-electricity.html
#AI #Anthropic

Zeiss SMO president Christoph Hensche stated that High-NA EUV is ready for mass production and orders are booked through 2027.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261019PD219/demand-production-president-2027-high-na-euv.html
#Semiconductors

Discrete language diffusion often uses absorbing corruption, leaving context fully trusted throughout generation.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22380
#MachineLearning

On-policy self-distillation enables large language models to learn from privileged information without a separate teacher model. We systematically evaluate 16 privileged information configurations across three multi-turn agentic benchmarks.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22512
#MachineLearning

Ridge alignment strengthens the relationship between model size and cross-dataset neighbourhood correspondence relative to native cosine geometry. Model scale improves linear recoverability more clearly than native cosine-neighbourhood correspondence.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22749
#MachineLearning

Researchers evaluate whether Large Language Models can predict the outcomes of machine learning experiments.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22599
#MachineLearning

Researchers introduce ΣΔ-Sampler, an optimizer that embeds a stochastic sampling-and-annealing step as a noisy quantizer inside a noise-shaping loop for binary and ternary AI models.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22621
#MachineLearning

Refusal is not robustness. We study hidden structure through bounded continuation analysis, which measures harmful reachability in autoregressive language models.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22754
#MachineLearning

Researchers introduced WebMix, a mixture hypernetwork that adapts web agents to unseen interfaces by composing previously trained low-rank adapters.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22539
#MachineLearning

Catastrophic forgetting is a challenge in supervised continual learning because adapting to new tasks causes representation drift and loss of previously learned information.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22439
#MachineLearning

Edit Discovery Optimization seeks to minimize missed edit opportunities while respecting an editor-cost budget. EDIT is an online algorithm for the partial-feedback setting.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.22661
#MachineLearning